DataTerminology.comNEW
Also known as: Step Size
The step size by which a convolutional filter advances over input data during convolution.
DESCRIPTION
Stride controls how a convolutional kernel traverses inputs, affecting output dimensions, receptive field coverage, and computational cost. Larger strides downsample feature maps, reducing spatial resolution and sometimes losing fine details. Smaller strides preserve spatial detail at higher compute and memory cost. Stride often interacts with padding, kernel size, and dilation to shape model capacity and efficiency. Tuning stride is crucial for CNN architecture design in vision and audio tasks.
ALTERNATIVE DEFINITIONS
Wikipedia
The stride is the number of pixels by which the filter moves across the input matrix.
Defines stride in the context of CNNs and spatial movement.
TensorFlow
Strides controls the sliding window for each dimension of input.
Emphasizes per-dimension stride parameterization in conv2d API.
PyTorch
Stride controls the stride of the convolution.
Minimal API-level description tied to Conv2d behavior.
RELATED TERMS
Convolution Kernel Size Padding Pooling Receptive Field
COMPETING TERMS
Dilated Convolution Transposed Convolution
KNOWLEDGE AREAS
TARGET AUDIENCE
AI Engineer Data Scientist MLOps Engineer Software Developer Hardware Engineer
EXTERNAL SOURCES
- Wikipedia - Convolutional neural network
- TensorFlow - conv2d API with strides
- PyTorch - Conv2d documentation
- Stanford CS231n - CNN notes discussing stride
TOP QUESTIONS
How does Stride affect output size in convolution?
Larger stride reduces output spatial dimensions by skipping positions. Smaller stride increases overlap, producing larger outputs and preserving detail.
Category: Example
Why is Stride relevant in CNN design?
Stride controls downsampling and computational cost, shaping receptive fields and model capacity for vision tasks.
Category: Relevancy
What are common mistakes when choosing Stride?
Using large strides early can discard fine-grained features. Ignoring padding interactions may cause undesired border artifacts.
Category: Mistakes
How does Stride interact with Padding?
With padding, stride determines coverage of border pixels. Without padding, larger strides shrink outputs quickly.
Category: Requirements
What is a competing parameter to Stride?
Dilated convolution expands receptive fields without downsampling, offering an alternative to increasing stride.
Category: Alternatives
When should Stride be greater than 1?
Use strides >1 for downsampling, reducing resolution and compute, often before pooling or in encoder stages.
Category: Opinion
What is the risk of using high Stride?
You may lose important spatial details and degrade accuracy on tasks requiring fine localization.
Category: Risk
How to tune Stride for performance?
Benchmark alternate stride, kernel, and padding combinations; profile FLOPs, memory, and accuracy to balance speed and quality.
Category: Management
Is Stride relevant outside images?
Yes, stride applies to audio, time series, and 1D convolutions for temporal downsampling.
Category: Overview